Bruker (Israel)

In recent years, Deep Learning (DL) algorithms employing artificial neural networks (NN) have been shown to be a powerful tool for the classification and quantification of images including intensity data from various x-ray scattering techniques. These algorithms generally require massive, labeled datasets for training and validating the NNs. Such datasets can be extremely difficult to obtain from measurements for several reasons. The size of the dataset is limited by the variation in available samples and the time available to measure the samples. The labels are limited by both the number and quality of available reference measurements. In this work we will discuss the use of DL for the analysis of critical dimension small angle x-ray scattering (CD-SAXS) data from high-aspect ratio (HAR) structures encountered in 3D NAND and DRAM memories. We have developed a novel solution for the automatic generation and labeling of large synthetic datasets using “realistic simulations” of the x-ray scattering data. Our solution includes instrumental artifacts such as background scattering and Poisson noise normally found in real x-ray images. Additionally, we have included structural variations based on parameters obtained from limited reference data, e.g., cross-section SEM images, which are critical to prevent overfitting of the DL model and improve the accuracy of the analysis. The realistic simulations are generated automatically using the NanoDiffract for XCD (NDX) software. The synthetic data are then used to train a convolutional neural network (CNN) that can be deployed and used for real-time inference. We demonstrate this approach by evaluating single-shot alignment and tilt of HAR memory hole structures. The CNN was validated with experimental data collected and analyzed using a Sirius-XCD® tool. Validation with reference data from production quality wafers resulted in an R² > 0.95 and precision 3σ < 0.02 degree with a significant reduction in measurement time per site.